This video introduces Kinematic-ICP, a novel approach designed to enhance LiDAR odometry for wheeled mobile robots operating on planar surfaces. It addresses the common problem of odometry drift and discrete jumps often observed in traditional wheel and LiDAR odometry systems, respectively, by integrating kinematic constraints directly into the Iterative Closest Point (ICP) optimization.
Kinematic-ICP leverages wheel odometry as an initial smooth guess and refines it with LiDAR measurements, optimizing for the robot's kinematic model inputs (e.g., linear and angular displacements) rather than a full 6D pose. This ensures a locally consistent and smooth estimated trajectory. The system also features an adaptive regularization mechanism to dynamically balance the influence of LiDAR and wheel odometry, improving robustness in challenging, feature-poor environments. Experimental evaluations on various robots and environments demonstrate that Kinematic-ICP achieves higher accuracy and smoother trajectories than state-of-the-art LiDAR odometry and fusion methods, while being computationally efficient.
Summary:
0:01 Introduction to Kinematic-ICP: Benedikt Mersch introduces Kinematic-ICP, a new method to improve LiDAR odometry for wheeled mobile robots by incorporating kinematic constraints.
0:10 The Need for Robot Odometry: Wheeled robots operating on flat surfaces (e.g., lab, warehouses) require precise estimation of their position and orientation changes over time (robot odometry), distinct from localization within a pre-built map.
0:45 Traditional Odometry Sensors: Explores the use of wheel encoders (for relative motion from wheel revolutions) and LiDAR (for pose estimation by aligning environmental measurements).
1:15 Limitations of Existing Methods:
Wheel Odometry: While smooth, it suffers from drift over longer distances (e.g., 200m test showed significant drift).
LiDAR Odometry (e.g., KISS-ICP): More globally consistent but exhibits undesirable "jumps" (e.g., 1-meter jump between scans, implying an unlikely 10 m/s velocity), violating the ROS REP-105 standard for smooth odometry.
2:41 The Goal: Smooth and Consistent Odometry: The objective is to efficiently estimate odometry that is locally consistent, smooth, and accurate to the true reference, without simply fusing disparate odometry sources.
2:58 Kinematic-ICP Framework:
Initial Guess: Starts with the smooth wheel odometry as a preliminary pose estimate.
LiDAR Correction with Kinematic Constraints: Refines this estimate using LiDAR scans by aligning them with a local map. Crucially, the correction (Ξ΄T) is constrained by the robot's kinematic model (e.g., a differential drive model for forward motion and rotation).
Optimization of Kinematic Inputs: Instead of optimizing for a full 6D pose, Kinematic-ICP optimizes for the control inputs (e.g., linear and angular displacements) of the kinematic model. This forces the resulting motion to be kinematically feasible and smooth.
4:12 Enhanced Odometry Results: Kinematic-ICP produces odometry that is both highly consistent with the ground truth and inherently smooth, effectively leveraging the strengths of both wheel odometry and LiDAR.
Key Takeaways (from video and paper details):
Accuracy & Smoothness: Significantly more accurate than wheel odometry and state-of-the-art LiDAR odometry systems like KISS-ICP, while maintaining local smoothness.
Efficiency: Processes data directly without fusing multiple odometry sources, leading to high computational efficiency (runs at 100 Hz on a single-core CPU, compared to 10 Hz for some fusion approaches).
Adaptive Regularization: Dynamically adjusts the weighting between LiDAR measurements and wheel odometry, improving robustness in challenging, feature-poor environments (e.g., long corridors).
Real-World Deployment: Successfully deployed in the Dexory fleet of robots in warehouses worldwide, demonstrating its practical applicability.
Open-Source: The project provides an open-source ROS 2 implementation for testing and development.